Opus 4.7 / 4.8 expose only 4 params — temperature, top_p, top_k, thinking.budget_tokens missing vs 4.5/4.6

[brunobuddy]

Summary

The two newest Opus entries (claude-opus-4-7, claude-opus-4-8) expose 4 parameters each, while claude-opus-4-5 and claude-opus-4-6 expose 8. The four missing knobs (temperature, top_p, top_k, thinking.budget_tokens) almost certainly still exist on the Anthropic API for these models — they appear to just not be backfilled.

This is more than a cosmetic gap: there's no signal in the payload that the data is incomplete, so a consumer (human or AI agent) reading claude-opus-4-8.json will confidently conclude the model has only 4 params. Missing data with no "partial" marker produces confidently wrong answers downstream.

Repro

for m in claude-opus-4-5-20251101 claude-opus-4-6 claude-opus-4-7 claude-opus-4-8; do
  echo "=== $m ==="
  curl -s "https://modelparams.dev/api/v1/models/anthropic/$m.json" \
    | python3 -c "import sys,json; d=json.load(sys.stdin); print(', '.join(p['path'] for p in d['params']))"
done

Output:

=== claude-opus-4-5-20251101 ===
max_tokens, temperature, top_p, top_k, thinking.type, thinking.budget_tokens, thinking.display, output_config.effort
=== claude-opus-4-6 ===
max_tokens, temperature, top_p, top_k, thinking.type, thinking.budget_tokens, thinking.display, output_config.effort
=== claude-opus-4-7 ===
max_tokens, thinking.type, thinking.display, output_config.effort
=== claude-opus-4-8 ===
max_tokens, thinking.type, thinking.display, output_config.effort

Expected

claude-opus-4-7 and claude-opus-4-8 should carry the same generation/sampling params as 4-5/4-6 (verified against Anthropic docs), unless Anthropic genuinely removed them — in which case that's worth an explicit note.

Suggested follow-ups

1. Backfill temperature, top_p, top_k, thinking.budget_tokens on claude-opus-4-7 and claude-opus-4-8 (and their -subscription variants).
2. Audit other recently-added models for the same regression.
3. Consider a completeness/status: "partial" marker (and/or a CI check that flags a new model exposing strictly fewer params than its predecessor in the same family) so incomplete entries are detectable rather than silently wrong.